Reading and changing are different risks
Most of what an agent does is reading: looking things up, opening pages, checking a list. If a read goes wrong you lose a little time and a little money, and nothing outside has changed. A change is different. A ticket that was filed, an email that was sent or a post that went out is acted on by someone else, often before anyone notices it was a mistake.
So the rule of thumb is simple: reads run at once, changes wait for a person.
| What the agent does | For example | Ask first? |
|---|---|---|
| Reading | Look up an order, list open tickets, read a file, open a web page | Usually at once. Nothing changes, and a wrong read costs only a little time. |
| Writing inside the workspace | Reply in a channel, save a note to knowledge | At once, within the agent’s daily limit. People see it and can correct it. |
| Changing something outside | Create a ticket, update a record, write a file on a computer | Ask first. Someone else will act on the result. |
| Speaking for you | Send an email, publish a post, reply to a customer | Ask first. It cannot be taken back. |
| Running something | Run a command, trigger an automation, start a payment | Ask first, every time. The effects are the hardest to predict. |
Why not approve everything, to be safe?
Because approvals that come too often stop being read. If every lookup needs a click, people learn to click without looking, and then the one request that matters gets the same automatic yes. An approval is worth having only while the person still reads it. Keep it for the actions that change something.
What a good approval request shows
- Who is asking. Which agent, in which conversation, so you know the context.
- What exactly will happen. The tool and every detail it will be sent, not a summary written by the agent. You approve the details, not the intention.
- What happens next. If you say yes, it runs with those details and the agent carries on. If you say no, nothing runs.
- A record. Who decided, and when, so the decision can be found later.
Who should approve
The person responsible for the agent, usually the one whose account pays for it. For a system the whole company shares, such as the ticket tracker or the shop, someone who looks after that system for the company should be able to decide as well. Nobody else.
When you are not sure, ask
A new tool appears, or a tool’s name does not say what it does. Treat it as a change until you know better: it costs one click to approve something harmless, and much more to undo something that was not.
How Agentic Crew does it
- When you connect a system through a connection, Agentic Crew looks at its tools. Those whose names suggest a change (create, send, update, delete and the like), and any it does not recognise, start on the “ask first” list. Lookups run at once. You can change either list.
- On a computer you have connected, an agent may read files in the folders you shared at once. Writing a file or running a command waits for approval every time, and that cannot be switched off.
- The request appears in the channel with the tool’s name and the exact details, with Approve and Decline. The agent’s owner decides; for a connection the workspace owners can decide too.
- An approved call is checked again before it runs: if the agent has lost the permission since it asked, it does not run. Then the agent continues with the result.
- Every decision is written to the workspace’s security log.
Approvals limit what an agent can change, not what it can spend. For money, set daily limits as well: see how to stop agents running up a bill.